In this work, we aimed at forecasting the number of SARS-CoV-2 hospitalized patients at 14 days to help anticipate the bed requirements of a large scale hospital using public data and electronic health records data. Previous attempts led to mitigated performance in this high-dimension setting; we introduce a novel approach to time series forecasting by providing an alternative to conventional methods to deal with high number of potential features of interest (409 predictors). We integrate Reservoir Computing (RC) with feature selection using a genetic algorithm (GA) to gather optimal non-linear combinations of inputs to improve prediction in sample-efficient context. We illustrate that the RC-GA combination exhibits excellent performance in forecasting SARS-CoV-2 hospitalizations. This approach outperformed the use of RC alone and other conventional methods: LSTM, Transformers, Elastic-Net, XGBoost. Notably, this work marks the pioneering use of RC (along with GA) in the realm of short and high-dimensional time series, positioning it as a competitive and innovative approach in comparison to standard methods.

International audience

In this article, we propose a novel architecture called hierarchical-task reservoir (HTR) suitable for real-time applications for which different levels of abstraction are available. We apply it to semantic role labeling (SRL) based on continuous speech recognition. Taking inspiration from the brain, this demonstrates the hierarchies of representations from perceptive to integrative areas, and we consider a hierarchy of four subtasks with increasing levels of abstraction (phone, word, part-of-speech (POS), and semantic role tags). These tasks are progressively learned by the layers of the HTR architecture. Interestingly, quantitative and qualitative results show that the hierarchical-task approach provides an advantage to improve the prediction. In particular, the qualitative results show that a shallow or a hierarchical reservoir, considered as baselines, does not produce estimations as good as the HTR model would. Moreover, we show that it is possible to further improve the accuracy of the model by designing skip connections and by considering word embedding (WE) in the internal representations. Overall, the HTR outperformed the other state-of-the-art reservoir-based approaches and it resulted in extremely efficient with respect to typical recurrent neural networks (RNNs) in deep learning (DL) [e.g., long short term memory (LSTMs)]. The HTR architecture is proposed as a step toward the modeling of online and hierarchical processes at work in the brain during language comprehension.

Understanding the mechanisms enabling children to learn rapidly word-to-meaning mapping through cross-situational learning in uncertain conditions is still a matter of debate. In particular, many models simply look at the word level, and not at the full sentence comprehension level. We present a model of language acquisition, applying cross-situational learning on Recurrent Neural Networks with the Reservoir Computing paradigm. Using the co-occurrences between words and visual perceptions, the model learns to ground a complex sentence, describing a scene involving different objects, into a perceptual representation space. The model processes sentences describing scenes it perceives simultaneously via a simulated vision module: sentences are inputs and simulated vision are target outputs of the RNN. Evaluations of the model show its capacity to extract the semantics of virtually hundred of thousands possible combinations of sentences (based on a context-free grammar); remarkably the model generalises only after a few hundred of partially described scenes via cross-situational learning. Furthermore, it handles polysemous and synonymous words, and deals with complex sentences where word order is crucial for understanding. Finally, further improvements of the model are discussed in order to reach proper reinforced and self-supervised learning schemes, with the goal to enable robots to acquire and ground language by them-selves (with no oracle supervision).

We propose a novel architecture called Hierarchical-Task Reservoir (HTR) suitable for real-time sentence parsing from continuous speech. Accordingly, we introduce a novel task that consists in performing anytime Part-of-Speech (POS) tagging from continuous speech. This HTR architecture is designed to address three sub-tasks (phone, word and POS tag estimation) with increasing levels of abstraction. These tasks are performed by the consecutive layers of the HTR architecture. Interestingly, the qualitative results show that the learning of sub-tasks en-forces low frequency dynamics (i.e. with longer timescales) in the more abstract layers. We compared HTR with a baseline hierarchical reservoir architecture (in which each layer is an ESN that addresses the same POS tag estimation). Moreover, we also performed a thorough experimental comparison with several architectural variants. Finally, the HTR obtained the best performance in all experimental comparisons. Overall, the proposed approach will be a useful tool for further studies regarding both the modeling of language comprehension in a neuroscience context and for real-time implementations in Human-Robot Interaction (HRI) context.

Gated working memory is defined as the capacity of holding arbitrary information at any time in order to be used at a later time. Based on electrophysiological recordings, several computational models have tackled the problem using dedicated and explicit mechanisms. We propose instead to consider an implicit mechanism based on a random recurrent neural network. We introduce a robust yet simple reservoir model of gated working memory with instantaneous updates. The model is able to store an arbitrary real value at random time over an extended period of time. The dynamics of the model is a line attractor that learns to exploit reentry and a nonlinearity during the training phase using only a few representative values. A deeper study of the model shows that there is actually a large range of hyperparameters for which the results hold (e.g., number of neurons, sparsity, global weight scaling) such that any large enough population, mixing excitatory and inhibitory neurons, can quickly learn to realize such gated working memory. In a nutshell, with a minimal set of hypotheses, we show that we can have a robust model of working memory. This suggests this property could be an implicit property of any random population, that can be acquired through learning. Furthermore, considering working memory to be a physically open but functionally closed system, we give account on some counterintuitive electrophysiological recordings.

The prefrontal cortex is known to be involved in many high-level cognitive functions, in particular, working memory. Here, we study to what extent a group of randomly connected units (namely an Echo State Network, ESN) can store and main-tain (as output) an arbitrary real value from a streamed input, i.e., can act as a sustained working memory unit. Furthermore, we explore to what extent such an architecture can take advantage of the stored value in order to produce non-linear computations. Comparison between different architectures (with and without feedback, with and without a working memory unit) shows that an explicit memory improves the performances.

In this paper we present a multi-modal human robot interaction architecture which is able to combine information coming from different sensory inputs, and can generate feedback for the user which helps to teach him/her implicitly how to interact with the robot. The system combines vision, speech and language with inference and feedback. The system environment consists of a Nao robot which has to learn objects situated on a table only by understanding absolute and relative object locations uttered by the user and afterwards points on a desired object to show what it has learned. The results of a user study and performance test show the usefulness of the feedback produced by the system and also justify the usage of the system in a real-world applications, as its classification accuracy of multi-modal input is around 80.8%. In the experiments, the system was able to detect inconsistent input coming from different sensory modules in all cases and could generate useful feedback for the user from this information.